Enhancing Tree Species Mapping in Arkansas' Forests through Machine Learning and Satellite Data Fusion: A Google Earth Engine-Based Approach

Abstract

Arkansas' subtropical climate nurtures extensive forested regions, particularly within the Ozark-
St. Francis and Ouachita National Forests. Despite this, the state lacks an up-to-date, high-
resolution map detailing the distribution of tree species within its forests. This study harnesses
the power of machine learning, specifically the Random Forest (RF), Gradient Boosting (GB),
Support Vector Machine (SVM), and K-Nearest Neighbors (K-NN) classifiers, within the
Google Earth Engine (GEE) framework. These classifiers are applied to classify the distribution
of tree species in Arkansas' forests by integrating data from various sources, including Sentinel-
1/-2, Landsat-8, and the National Agriculture Imagery Program (NAIP). The study evaluates the
classification accuracy of single-sensor images against fused composites, revealing that the fused
Landsat-8 and Sentinel-1 data achieve the highest validation accuracy at 0.8875. This is closely
followed by single-sensor Sentinel-1 and Landsat-8, which yield validation accuracies of 0.8863
and 0.8859, respectively. Among the classifiers, RF demonstrates the highest accuracy, followed
by GB, SVM, and K-NN when applied to fused Landsat-8 and Sentinel-1 images. This study
incorporates the Shapley Additive Explanations (SHAP) to elucidate feature importance and
introduces a weighted ensemble method, resulting in a remarkably accurate tree species
distribution map with an accuracy score of 0.9772. This research highlights the efficacy of
combining machine learning algorithms and fusing satellite images to significantly enhance tree
species classification accuracy. Moreover, the study capitalizes on explainable AI (XAI)
principles and leverages the cloud computing capabilities of GEE to create a more precise, high-
resolution tree cover map on a regional scale.